Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design

Meta's engineering team says the new version of Andromeda, its system for picking which ads might be shown to someone, now scores every eligible ad at near ranking-level quality and surfaces top candidates at over one billion ads per second. Meta calls that a 100x lift over the 2024 system. Since 2025, Meta says Andromeda has delivered a 2.4% increase in conversions on Facebook.

The pressure behind it is telling. Generative AI tools have helped advertisers create hundreds of millions of new ads, swelling the pool Meta has to sort through. AI is now generating ads faster than older systems could evaluate them.

As pure engineering, this is impressive and worth studying. Meta borrowed mixture-of-experts ideas from large language models, used specialized user models as experts, and claims near ranking-level quality at roughly 300x fewer inference-time FLOPs than the 2024 model. It then fused model and system work into a single GPU kernel and moved index matching and selection onto the same GPU server, which it says cut network traffic about 100x. The lesson is that model, system and hardware should be designed together.

We'd also say plainly what this is for. The goal is more precise ad matching at lower cost. Meta lists the inputs: sequences of ad engagement, cohort representations, LLM-derived semantic features and engagement across Meta's surfaces. Every efficiency gain here is an efficiency gain in modeling people's interests for commercial persuasion.

There's a fair case for users. More relevant ads beat irrelevant ones, and Meta frames it that way. But the feedback loop is what I'm watching. Generative tools flood the system with ads, a faster retrieval engine sorts them, and the whole thing gets better at finding the version that converts. If that scales unchecked, the harm is subtle: more optimized persuasion, less visibility into why a given ad found you.

Meta also says the pipeline is not at its ceiling, with INT8 quantization and Grace Blackwell work still ahead. Expect this machine to keep getting stronger.

Retrieval is the first stage of Meta's ad recommendation process, narrowing a huge pool of eligible ads to a small set in milliseconds before a costlier ranking stage scores them. The 2024 Andromeda launched on NVIDIA's GH200 Grace Hopper Superchip. In early 2026 Meta and NVIDIA began optimizing each pipeline component together, reporting 2x to 5x speedups per component.